The Taub Faculty of Computer Science Events and Talks

In image restoration, traditional supervised methods that seek to restore the source enjoy exceptional distortion performance but lack visual quality. With the emergence of powerful generative algorithms, many approaches focus on image realism and diversity, but forsake faithfulness to the source. Motivated by recent theoretical findings, we present a practical algorithm that optimizes source fidelity while aiming for photo-realistic results. Our method optimally transports the distribution of MMSE estimate to the natural image distribution using a simple patch-level deep representation (as simple as an auto-encoder). The results of our experiments demonstrate that we can effectively improve the perceptual quality of MMSE estimates on severe degradations.